arXiv:1604.00322v1 [cs.DS] 1 Apr 2016 ⋆⋆ ie olcin fe hs rbesaesuidwt epc to respect with studied subcollection are weight problems maximum the these find Often to collection. want given we a problem a In Introduction 1 acigi h aeweealcpcte r nt The unit. are capacities all where case the is matching aaiylimits. capacity htfreeyvertex every for that pca ae like cases special nrdcdfrgah n[6.Frti rbe aheg sgvnade a given is edge each problem this For [26]. in graphs for introduced yegah”;mthn splnma-ieslal for solvable polynomial-time is matching ”); s nytenieL eaain.Tewrtcs ai ewe th between ratio the worst-case The called relaxations. is LP solution naive the only use eeta most at select l- etrw eoe the recover we vector all-1 oyoiltm)a nerlslto fvlea es 1 least at value of solution integral an time) polynomial ntnecnalwec set each allow can instance euiyAmnsrto ne otatDE-AC04-94AL8500 fo contract Corporation, under Martin laborator Administration Lockheed multi-program Security of a subsidiary owned is wholly Laboratories National Sandia n forrslscniestegnrlzto of generalization the considers results our of One yegahcmthn rbesaeotnsuidvalna progra linear via studied often are problems matching Hypergraphic The ilacin nwihayn a i tmost at win can anyone boun which upper in car auctions gap more integrality rial algorithmic a bipartite using the obtained Using are results These method. LP. natural the to iatt,w r bet mrv h prxmto ai to ratio approximation the previously-known improve a to matches able it are since we power, bipartite, prime a than more na most at in subcollect maximum-weight in a seek generally, we and weight, associated Abstract. prxmto loih for algorithm approximation advle.Tebs nw prxmto loih o thi for algorithm approximation known best The values. mand egv e loih,bsdo oa ai,ta ban th obtains that ratio, local on based way. algorithm, simpler new a give We uto that auction e prxmtosfragnrlzto ftercnl int recently of generalization the the of consider generalization also We a for approximations new b eeaie yegahMthn i trtdPacking Iterated via Matching Generalized mthn rbe eeaie acig h nu pcfisalimit a specifies input the matching: generalizes problem -matching b v b -iesoa matching 3-dimensional escnann each containing sets In k t eso h ucleto.Ormi euti linear-progr a is result main Our subcollection. the of sets apoiaeymxmzssca efr.W loso that show also We welfare. social maximizes -approximately -hypergraph k hprrp acig eaegvnacleto fst fsi of sets of collection a given are we matching, -hypergraph v nerlt gap integrality h u fthe of sum the , 2 1 nvriyo otenClfri,LsAglsC 08,US 90089, CA Angeles Los California, Southern of University adaNtoa aoaois luuru M815 USA 87185, NM Albuquerque Laboratories, National Sandia b mthn problem. -matching e ob eetdmlil ie pt oeupper some to up times multiple selected be to b mthn,isedo ijitesw eur hteeyel every that require we disjointness of instead -matching, k -hypergraph jsParekh Ojas An . v riaymthn eut when results matching ordinary ; d n oa Ratio Local and b vle fteicdn de hudb tmost at be should edges incident the of -values mthn to -matching [18]. LP-relative [email protected] [email protected] b mthn.Ti ete h nerlt a when gap integrality the settles This -matching. 1 t tm,teei rtfli-xetto polynomial-tim truthful-in-expectation a is there items, n ai Pritchard David and eadmatching demand α b uncapacitated mthn to -matching /α apoiainalgorithm -approximation aae n prtdb adaCroain a Corporation, Sandia by operated and managed y 0. k h ..Dprmn fEeg’ ainlNuclear National Energy’s of Department U.S. the r ie h Psotmlvle—ti ohupper both this — value optimal LP’s the times ,w hwta o h aiyo combinato- of family the for that show we d, ,wiei is it while 2, = oue one-oo acigproblem. matching bounded-color roduced o hs esaepiws ijit More disjoint. pairwise are sets whose ion k rbe a ai 2 ratio has problem s − flapiaino the of application eful aeapoiainrtoi much a in ratio approximation same e oe on.We h yegahis hypergraph the When bound. lower ,wihi lobs osberelative possible best also is which 1, Potmmadteotmlintegral optimal the and optimum LP e eadmatching demand aei where is case hr de aennnfr de- nonuniform have edges where , advalue mand h aiu e size set maximum the 2 mn eaain.I hspprwe paper this In relaxations. mming b migbsd( based amming steal1vco.A vector. all-1 the is ea most at ze b fpiws ijitst within sets disjoint pairwise of u eut ietyimply directly results our v APX o vr etx n ecan we and vertex, every for capacity ⋆⋆ A, hr for -hard d k soeta rdcs(in produces that one is e , c on n enwconstrain now we and , = trtdpacking iterated k k oinoriginally notion a , mn appears ement ∞ − → ahwt an with each , -hypergraphs. limit b k ..teeaeno are there i.e. , v When . − k k k + 1 c ,ee in even 3, = e sone is ie n“ on (i.e. . b -matching Simple k 1 )- d e sthe is k b - - bounds the integrality gap by α and gives an α-approximation algorithm. Many classical approximation algorithms are LP-relative; so the notion is not novel, rather, this terminology helps us be concise.

1.1 Results Our main result is the following theorem. 1 Theorem 1. There is an LP-relative (k − 1+ k )-approximation algorithm for k-hypergraph b-matching, for any capacities. In [23] one of the authors announced, without a proof, a weaker result than the above theorem, namely an 1 upper bound of k − 1+ k on the integrality gap. Here we give an algorithm to find an integral solution matching this bound in polynomial time, requiring a significant extension of the techniques presented in [23]. 1 For the special case b = 1, F¨uredi, Kahn and Seymour [15] proved an upper bound of k − 1+ k in 1993, while Chan & Lau [10] recently gave the first polynomial-time algorithm matching this bound. Their technique does not directly extend to the k-hypergraph b-matching case. The technique that we use to prove Theorem 1 is iterated packing, the same technique from [23]. Part of the contribution of the present paper is to simplify and extend some of the the approaches from [23] and [10]. Our main technical innovation is, using iterated packing, to explicitly specify particular additional solutions as ineligible for packing: not only solutions that would be ineligible for the original problem, rather we additionally prohibit solutions exceeding the ceiling of the current fractional solution. Theorem 1 is tight for infinitely many k: when k − 1 is a prime power, as observed in [15], the projective PG 1 plane (2, k − 1) of order k − 1 yields a matching lower bound of k − 1+ k on the integrality gap. It is an interesting open question to settle the integrality gap for any other values of k. We are able to determine the exact integrality gap for another interesting class of hypergraphs. Call a hypergraph bipartite ([1]; cf. [24]) if, for some distinguished subset U of vertices, every hyperedge contains exactly one vertex from U. Theorem 2. There is an LP-relative (k−1)-approximation algorithm for bipartite k-hypergraph b-matching, for any capacities. Chan and Lau [10] proved Theorem 2 in the special case that b = 1 and the instance is k-dimensional3. Proving Theorem 2 is similar to Theorem 1 plus extending an observation of [10] from k-dimensional hypergraphs to bipartite ones. Like Theorem 1, a matching integrality gap lower bound is known [14, p. 157] when k − 1 is a prime power: the hypergraphic dual of the affine geometry AG(2, k − 1), i.e. a truncated , has integrality gap k − 1. We obtain the following interesting corollaries from the bipartite case. In the bounded-color k-hypergraph b-matching problem we are given an instance of the k-hypergraph b-matching problem along with a partition of the edge set into l color classes, E = E1 ∪···∪ El, and a positive integer wi for 1 ≤ i ≤ l. We seek a feasible k-hypergraph b-matching of maximum weight such that at most wi edges from class Ei are selected for each i. Corollary 1. There is an LP-relative k-approximation for bounded-color k-hypergraph b-matching. Corollary 2. For combinatorial auctions where each bidder can win at most (k−1) items, there is a random- ized polynomial-time mechanism that, in expectation, is both truthful and (k − 1)-approximately maximizes social welfare. We are not aware of any prior results for this extremely natural class of combinatorial auctions, cf. [22, Ch. 12]. The proof of Corollary 2 uses the mechanism of Lavi and Swamy [21], where the distinguished vertices in the bipartite hypergraph correspond to the bidders. For this application, it is crucial that Theorem 2 gives an LP-relative approximation in polynomial time. Finally, we give a new short proof of the following known theorem: 3 A hypergraph is k-dimensional if for some k-partition of the ground set, every edge intersects every part exactly once.

2 Theorem 3 ([23]). There is an LP-relative 2k-approximation for k-hypergraph demand matching.

Our simpler proof is based on the local ratio method, rather than the iterated packing used in [23]. We rely on a connection in [5, p. 12] between local ratio and iterated packing.

1.2 Related Work

As Tutte observed [28], both in edge-weighted graphs and in the cardinality case, uncapacitated graphic b- matching can be reduced to matching by replacing each vertex by bv clones. Each edge uv is likewise cloned bubv times. This reduction has two problems: (1) the clones cause an exponential increase in the instance size (from lgkbk1 to kbk1); and (2) it does not work in the capacitated case, since we need to prevent too many clones of the same edge from being selected. Cloning applies to hypergraphs, too, but has the same two problems. Algorithmically, we can often avoid (1) by not dealing with the clones explicity. For graphs we can fix problem (2): an edge-trisecting reduction [28] (see also [25, p. 562]) extends cloning to work on capacitated instances. But for hypergraphs, there is no known workaround for problem (2). As a strawman, let us mention that one can reduce capacitated b-matching in k-hypergraphs to unca- pacitated b-matching in (k + 1)-hypergraphs, by inserting new vertices in each hyperedge and by moving each edge’s capacity to the b value of its new vertex. One can even then apply cloning. But this is not that useful for us: e.g., we cannot use the previously-known b = 1 case of version Theorem 1 to even prove the nonconstructive version of Theorem 1 for general b, since this reduction increases the hyperedge size from k to k + 1. Algorithmically, the simple (capacity c = 1) case of b-matching is the hardest. The proof is standard, by fixing the integer part of an optimal fractional solution.

Observation 1 Given an (LP-relative) α-approximation to simple b-matching in k-hypergraphs, we can obtain the same quality of approximation for general capacities.

Hypergraph matching. Matching problems in k-uniform hypergraphs are well-studied algorithmically. For k any fixed ε> 0 the best known approximation ratios are 2 + ε for the unweighted version by Hurkens and k+1 Schrijver [17] and 2 + ε for the weighted version by Berman [8]. In the case k = 3, the algorithmic results of [10] give an ε-improved approximation ratio of 2 for 3-dimensional matching. On the other hand, Hazan, k Safra and Schwartz [16] showed that the problem is hard to approximate within a factor of Ω( log k ) unless P = NP, even in the k-dimensional case.

Hypergraph b-matching. For b-matching in k-hypergraphs, Krysta [20] gave a greedy k + 1-approximation for the simple case, and Young & Koufogiannakis [19] gave a k-approximation for the uncapacitated version. Both of these approximation algorithms give LP-relative guarantees. An improvement in some cases was recently obtained by the k-exchange system framework of Feldman et al. [13]. The b-matchings form a k- exchange system (this is explicit only for k = 2 in [13]). In this way one can obtain a local search-based k+1 ( 2 + ε)-approximation algorithm for weighted k-hypergraph b-matching. However, its running time is exponential in k and it does not give any LP-relative guarantee. It may be tempting to think that the b-matching problem in hypergraphs is a simple extension of 1- matching in hypergraphs because the theory and algorithms for b-matching in graphs closely relate to those for 1-matchings. As evidenced by the results above, this does not appear to be the case. An approximation algorithm that runs in time polynomial in k with guarantee better than k for k-hypergraph b-matching had been an open problem that we resolve with this work. Our methods are LP-based, whereas local search seems to give the best known results; however, the bounding techinques used in local search for hypergraph 1- matching do not seem to readily extend to the hypergraph b-matching case. For example, Arkin and Hassin [2] give a local search (k − 1+ ε)-algorithm for weighted k-hypergraph 1-matching; however, as a warmup they present a trivial bound of k — even this trivial bound does not easily extend to the k-hypergraph b-matching case.

3 Other work. Pseudo-greedy methods similar to iterated packing have been successfully applied to several packing and coloring problems, including multicommodity flows on trees [11], independent sets in t-interval graphs [5], and weighted of bipartite graphs [12]. Iterated packing is a means of obtaining an approximate convex decomposition; Carr and Vempala [9] have shown a strong connection between the latter and approximation ratios of LP-based approximation algorithms. As mentioned earlier, a 2k-approximation for k-hypergraph demand matching is known [23]; a better ratio of 3 is possible when k = 2 [23]. These nearly match (exactly match, when k = 2) the best known lower bound of 2k − 1 [3] on the integrality gap of the natural LP relaxation (this construction does not require that k − 1 is a prime power). Bansal et al. [3] devised a deterministic 8k-approximation and a randomized (ek + o(k))-approximation for the more general problem of approximating k-column-sparse packing integer programs. Stamoulis very recently introduced the bounded-color matching problem (defined above in the more general hypergraph context) and devised a 2-approximation [27]. This result is also based on iterated packing. Stamoulis observes that the bounded-color matching problem is a special case of 3-hypergraph b-matching. 1 In fact it is suggested in this paper that a polynomial-time (k − 1+ k )-approximation for k-hypergraph b- matching may be possible. Our work was developed independently, and we observe that our results generalize Stamoulis’s results, since the special hypergraph b-matching instances obtained by the reduction he suggests are bipartite, and we are able to leverage Theorem 2 to give a k-approximation for the more general bounded- color k-hypergraph b-matching problem, which we introduce here. We will exploit the interplay between LP-relative approximation algorithms and convex decompositions — an equivalence between the two was shown by Carr & Vempala [9]. The Lavi-Swamy [21] mechanism combines techniques from [9] with the VCG mechanism. We give an overview of iterated packing in the next section. There, we also introduce a structure theorem from [10] and its specialization to bipartite instances, versions of which will be used throughout the paper. Next, to further introduce the iterated packing methodology, we give an iterated packing proof of the same result, although it does not run in polynomial time. This is extended to b-matching in Section 4, which contains our main technical innovations. First an existential proof is given (Algorithm 2) and then finally Algorithm 3 proves Theorems 1 and 2 constructively. Then in Section 7 we present the proof of Theorem 3, which is based on the local ratio method.

2 Iterated Packing Overview

The notion of an approximate convex decomposition is essential to iterated packing, as the latter iteratively builds such a decomposition for a given fractional solution. Here we present a slightly different notion of an approximate convex decomposition than usually considered.

Definition 1. For α ≥ 1, define α-convex multipliers to be any collection of nonnegative reals whose sum is i α. Likewise, we say that x is an α-convex combination of the points {x }i if there are α-convex multipliers i {λi}i so that x = i λix . The utility of α-convexP combinations is that they provide a convenient way to talk about integrality gaps without rescaling as was done in [9] or [23]. Proposition 1 ([9]). If every feasible LP solution for a packing program can be written as an α-convex combination of integral feasible solutions, then its integrality gap is at most α.

Proof. We need to show that for any nonnegative weight function w, if x∗ is the fractional solution that maximizes w(x∗), then there is an integral solution of weight at least w(x∗)/α. A random solution from ∗ i the α-convex combination representation of x , drawing x with probability λi/α, has expected weight λi i ∗ i i α w(x )= w(x )/α. So one of the x has at least this weight. P(In fact [9] also proves an algorithmic converse, used also by the Lavi-Swamy framework [21] underlying Corollary 2.)

4 We will use Proposition 1 as follows: we develop a polynomial-time algorithm to write fractional hyper- graph b-matchings as ρ-convex combinations of feasible integral b-matchings. Then by Proposition 1, we get the LP-relative ρ-approximation algorithm claimed in Theorems 1 and 2. In [23] the idea of iterated packing was introduced. Each iteration, called a packing step, updates the current α-convex combination to a new one, increasing some terms of the combination on one coordinate. i i Definition 2 (Packing step). Let us be given an α-convex combination x = i λix where the x are feasible integral solutions, an edge e to pack, and a target value t ∈ R+. We may think of a packing step as packing the edge e into some of the solutions xi such that each resulting solution isP still feasible and that we have packed e into solutions with a total mass of t, i.e. the sum of corresponding λi is t. E Let χe be the vector in R with coordinate 1 on e and 0 elsewhere. A packing step will replace some ′ i i ′ i 0 ≤ λi ≤ λi portion of each x with x + χe, where we allow λi > 0 only when x + χe is feasible. Therefore ′ i ′ i i(λi −λi)·x + i λi ·(x +χe), the result of the packing step, expresses x+tχe as an α-convex combination of integer feasible solutions. P P For a packing step to actually be feasible, it is clearly both necessary and sufficient that the set P = {i | i x + χe feasible} of solutions into which e can be packed must satisfy λ(P ) ≥ t. For the sake of polynomial-time implementation of our final algorithm, note we can ensure at most one i ′ has λi ∈{/ 0, λi} in the above argument, so that each packing step increases the number of terms by at most one. Alternatively we could use Carath´eodory’s theorem which guarantees that any α-convex combination can be rewritten as one with at most d + 1 terms where d is the number of coordinates. The basic iterated packing formula starts with a fractional solution x in hand and iteratively constructs an integral solution by starting with an empty hypergraph on V . The edges are processed in some order, and for each edge e, a packing step is performed on e with a target value of xe. One key fact about iterated packing is that when a target value is larger, packing is easier, hence iterated packing shows how large fractional values facilitate approximation for packing problems much like iterated packing does for covering problems. The basic approach may be refined in several directions. One may start with base integral solution that is non-empty hypergraph. This was explored in [23] to derived an improved approximation for the demand matching problem. Another improvement is to consider a specific ordering of edges. This key idea driving our algorithm is analyzing an ordering of edges which allows us to obtain a polynomial-time algorithm. Although, as announced in [23], extensions of ideas from [23] may be used to derive an upper bound of k − 1+1/k on the integrality gap for the k-hypergraph b-matching problem, the bound is non-constructive and does not give a polynomial-time algorithm. We show that by considering an ordering of edges that was first studied by Chan and Lau [10], we obtain a polynomial-time k − 1+1/k- approximation. This ordering is based on vertices of small degree in an extreme point solution, which in turns allows one to argue that there is an edge with large fractional value. The lemma below shows that we can find a vertex of sufficiently small degree. Let {Av,e}v,e be the 0-1 incidence matrix for our k-hypergraph: it has rows for vertices and columns for edges, with at most k ones per column. When x∗ is an extreme point solution to the matching LP ∗ {0 ≤ x ≤ 1 | Ax ≤ 1}, elementary properties of polyhedra show that the incidence matrix of {e | 0 < xe < 1} has linearly independent columns. This makes the following lemma useful: it was proven by Chan and Lau for the general case, while the bipartite case follows from generalizing their arguments about the k-dimensional case. Lemma 1. If the incidence vectors of ∅ 6= E′ ⊆ E are linearly independent, then some vertex in (V, E′) has degree between 1 and k. In the bipartite case, the upper bound can be strengthened to k − 1. Proof. The first part is a counting argument. The incidence matrix retains its rank if we delete the all-zero rows, leaving only those rows corresponding to the set V ′ of vertices with nonzero degree. The number of such vertices must satisfy |V ′| ≥ |E′| or else rank |E′| could not be achieved. Since each column has at most k unit entries, there are at most k|E′| unit entries in the whole matrix. So averaging, some row has at most k|E′|/|V ′|≤ k nonzeroes, and this gives the desired vertex. For bipartite hypergraphs, examine the situation in which equality holds. This can only happen if |E′| = |V ′| and the matrix has exactly k ones per row and per column. Let U be the subset of vertices so that every

5 hyperedge intersects U exactly once. So, each hyperedge intersects the complement of U exactly k − 1 times. Therefore, the vector in RV with (−k − 1) entries in U and unit entries elsewhere is orthogonal to all rows, contradicting that the adjacency matrix has full rank.

In order to talk about both the general and bipartite cases in a unified way, define

k − 1+ 1 in the general case, and ρ := k (k − 1 in the bipartite case.

Additionally, define the degree bound

k in the general case, and µ := (k − 1 in the bipartite case.

3 Non-Polynomial Time Algorithm for k-Hypergraph Matching

1 We now give an alternate proof that k-hypergraph matching has integrality gap of at most k +1 − k . The algorithm behind this proof does not run in polynomial time. However, this section also introduces the notation and steps involved in iterated packing, which we will extend in the next section to get our main result.

Lemma 2 ([23]). In k-hypergraph matching, a packing step to bring x to x+tχe, where x+tχe is a feasible fractional solution, is possible if α ≥ k − (k − 1)t.

i Proof. Let Qv, for each v ∈ e, be the set of solutions i for which x +χe is not feasible. We have λ(Qv) ≤ 1−t 4 since x + tχe is feasible . We need room (disjoint in the worst case) for all such Qv, plus an additional t to pack the new edge in solutions that permit it, giving the bound k(1 − t)+ t = k − (k − 1)t.

We can indeed get large coordinates using the following strengthening of Lemma 1. Lemma 3. Any nonzero extreme point solution x to the k-hypergraph matching polytope has some fractional coordinate at least 1/µ.

Proof. If xe = 1 for any coordinate then we are done, so suppose otherwise. We know from elementary linear algebra that there is a set V ′′ of vertices and a set E′′ of edges so that x is the unique solution to ′′ ′′ xe =0, ∀e∈ / E ; x(δ(v)) =1, ∀v ∈ V . Then the same counting argument as in Lemma 1 (resp. and the same linear independence in the bipartite case) ensures that some v ∈ V ′′ is incident on at most k (resp. k − 1) edges. Since it has x(δ(v)) = 1 the e ∈ δ(v) maximizing xe satisfies the lemma.

Using this, we obtain an iterated packing algorithm for the k-hypergraph matching problem, which is displayed as Algorithm 1. Note that this algorithm is presented as a recursive top-down variant of iterated packing, while the basic version in the previous section was presented as a bottom-up algorithm for ease of exposition. Another more crucial deviation of this algorithm from the basic iterated packing formula is that since our analysis requires an extreme point, we must express each non-extreme solutions as convex combinations of extreme points, and we use that:

A convex combination of α-convex combinations is an α-convex combination. (1)

In fact this is the reason the algorithm is not guaranteed to run in polynomial time; however, the algorithm does terminate since the number of nonzero coordinates of x decreases in each recursive call.

4 i In detail, the solutions x for i ∈ Qv have degree 1 at v, so by the definition of a convex combination (Ax)v = λ(Qv), but (Ax)v ≤ 1 − t since, by feasibility, 1 ≥ A(x + tχe)v =(Ax)v + t.

6 Algorithm 1 HM∗(V,E,x) // write x as ρ-convex comb. of 0-1 solutions

1: If x = 0 return the trivial ρ-convex combination λ1 = ρ,x1 = 0. E 2: If x is not an extreme point solution to {x ∈ R+ | Ax ≤ 1}, 3: Write x as a convex combination of extreme point solutions. 4: Recurse on each extreme point and return their result combined via (1). ′ ′ 5: Pick e so that xe is maximized and let x be x except with xe set to zero. i ∗ ′ 6: Recurse: (x ,λi)i := HM (V,E,x ). i 7: Packing step: pack xe of e into (x ,λi)i and return the result.

Proposition 2. Given any LP solution x, Algorithm 1 returns an expression of x as a ρ-convex combination of integral solutions.

1 Proof. This follows from Lemmas 2 and 3, since if µ = k we have k − (k − 1)/µ = k − 1+ k = ρ, and if µ = k − 1 we have k − (k − 1)/µ = k − 1= ρ.

This completes the non-polynomial time iterated packing proof that the integrality gap for matching is at most ρ. Next, we extend it to b-matching.

4 Iterated Packing and k-Hypergraph b-Matching

In this section, which contains the main new iterated packing technique, we build on the ideas from the previous section. We begin with a non-constructive iterated packing algorithm to show that the integrality gap for k-hypergraph b-matching is at most ρ. Then, we move to a constructive version via iterated packing that runs in polynomial time. By Observation 1, we assume unit capacities (simple b-matching). We will use the following statement, whose proof is analogous to Lemma 3. Lemma 4. Any nonzero extreme point solution x to the k-hypergraph b-matching polytope has some frac- tional coordinate at least 1/µ. The naive adaptation of iterated packing (Algorithm 1) to b-matching would involve writing the input as a convex combination of extreme point solutions to {x ∈ [0, 1]E | Ax ≤ b}, working with α-convex combinations of integer 0-1 solutions to Ax ≤ b. However, this approach is unworkable. When we try to mimic Lemma 2, as b gets larger, we cannot bound λ(Qv) by anything less than 1, giving an approximation ratio of k or worse. To fix this problem, we will enforce two additional conditions. One of these conditions, the main driver of the new proof, is that the strengthened degree bound Axi ≤⌈Ax⌉ must hold in every level of the recursion (rather than the unworkable requirement that solutions merely respect the final target degrees). The second condition is that the λ-mass of solutions meeting this strengthened bound with equality cannot be more than h(Ax)v i (here h·i denotes the fractional part), except in the degenerate case that (Ax)v is integral. Intuitively (i) balances the number of edges packed at a vertex across the solutions xi, avoiding the trouble that the naive approach would encounter in future iterations, while (ii) helps achieve (i) inductively. A modified packing step is a packing step that, given a solution (x, λ) satisfying both of these properties, produces another (x′, λ′) satisfying both of these properties. Then the definition of the resulting algorithm, Algorithm 2, is as follows. We will prove by induction that the algorithm succeeds in finding packings meeting both conditions. Lemma 5. For any 0 ≤ x ≤ 1, HbM∗(V,E,x) returns an expression of x as a ρ-convex combination i i of 0-1 x that satisfies (i) Ax ≤ ⌈Ax⌉ for each i, and (ii) for every v such that (Ax)v is non-integral, i λ({i | (Ax )v = ⌈Axv⌉}) ≤ h(Ax)v i.

For the proof, it is helpful to realize that we use (Ax)v interchangeably as x(δ(v)), and that it represents the “degree” of x at v.

7 Algorithm 2 HbM∗(V,E,x) // write x as ρ-convex comb. of special 0-1 solutions

1: If x = 0 return the trivial ρ-convex combination λ1 = ρ,x1 = 0. // as before E 2: If x is not an extreme point solution to {y ∈ [0, 1] | Ay ≤ ⌈Ax⌉}, 3: Write x as a convex combination of extreme point solutions. // as before 4: Recurse on each extreme point; return their combination via (1). // as before ′ ′ 5: Pick e so that xe is maximized; let x be x with xe set to zero. // as before i ∗ ′ 6: Recurse: (x ,λi)i := HbM (V,E,x ). // as before i 7: Modified packing step: pack xe of e into (x ,λi)i and return the result.

Proof. The base case and the non-extreme case are easy; while the extreme points decomposing a non- extreme solution may have smaller values for ⌈Ax⌉, this does not hurt us. So we only need to deal with the case that x is extreme and nonzero, where e is chosen with xe ≥ 1/µ. To prove that the modified packing step can always be carried out while satisfying (i) and (ii), we again bound a set of unpackable solutions. Specifically, our goal will be to define sets Qv for each v ∈ e such that any packing step that avoids adding e to any of the solutions v∈e Qv will satisfy (i) and (ii) for x, and such that the sets Qv are λ-small enough that e always has room to be added. S ′ For each v ∈ e, there are three cases, the main distinction being whether ⌈(Ax )v⌉ = ⌈(Ax)v⌉. Note that these terms are either equal, or differ by one.

′ – Case (I), (Ax )v = 0. This packing is trivial, set Qv = ∅. ′ – Case (II), ⌈(Ax)v ⌉ = ⌈(Ax )v⌉ 6= 0. Proving (ii) is vacuous when (Ax)v is integral, and otherwise it ′ follows easily by induction since h(Ax)v i = h(Ax )vi + xe and at most xe of λ-mass of solutions will have its degree increased at v. To show (i) is satisfied inductively, just like in Section 3, define Qv to be the i i set of i with (Ax )v = ⌈(Ax)v⌉; e can be added to any other x without violating the degree constraint. ′ The terms (Ax)v and (Ax )v differ by xe and have the same integer ceiling, so by induction on (ii), we ′ have the bound λ(Qv) ≤ hx (δ(v))i ≤ 1 − xe showing that Qv is not too big. This bound will be used later. ′ ′ i – Case (III), ⌈(Ax)v ⌉ = 1+ ⌈(Ax )v⌉ and (Ax )v 6= 0. Then satisfying (i) at v is easy (since all x have ′ degree at most ⌈(Ax )v⌉ at v) but we must design Qv so that (ii) is satisfied after the packing step. If (Ax)v is integral any packing works (we can take Qv = ∅), so assume the opposite. Moreover, when ′ i (Ax )v is integral, by (i) all solutions x have degree less than ⌈(Ax)v ⌉ at each v ∈ e, and since we are only packing xe = h(Ax)v i amount of e, (ii) is also satisfied by any possible packing. ′ Hence, assume both (Ax )v and (Ax)v are non-integral. If we pack e arbitrarily, the total weight of new solutions with degree ⌈(Ax)v⌉ at v could be too large to satisfy (ii). Therefore, we will define Qv ′ i ′ to exclude some subset of the solutions Qv := {i | (Ax )v = ⌈(Ax )v⌉} that could rise to have this ′ ′ ′ degree. We have λ(Qv) ≤ h(Ax )vi from (ii) inductively. We now define Qv to be some subset of Qv with ′ ′ λ(Qv)=1 − xe. This is not possible if λ(Qv) < 1 − xe but in this case we just define Qv := Qv. Also, ′ 5 even if no subset of Qv has λ-value exactly 1 − xe we can split a term of the ρ-convex decomposition to achieve this. The point of this Qv is that, using (ii) inductively, the post-packing total λ-value of the ′ ′ solutions with degree ⌈(Ax)v ⌉ at v will be at most λ(Qv \ Qv) ≤ h(Ax )vi− (1 − xe)= h(Ax)v i; the latter ′ equality holds since (Ax)v = (Ax )v + xe and by the hypotheses of this case. So these Qv allow us to inductively satisfy (i) and (ii), on top of which λ(Qv) ≤ 1 − xe.

In all cases, λ(Qv) ≤ 1 − xe. Analogous to Lemma 2 there is enough room to complete the packing step so long as ρ ≥ xe + λ( v∈e Qv). By a union bound this would be implied by ρ ≥ xe + k(1 − xe). This gives the same analysis as before (Proposition 2) in terms of our bounds on xe and ρ, so the modified packing step succeeds and we areS done.

i i i i i 5 Splitting means to replace the term (x ,λ ) with two terms (x ,p), (x ,λ − p) with distributed λ-mass on the same i integer solution x .

8 4.1 Polynomial-Time Iterated Packing for k-Hypergraph b-Matching

Finally, we give our main algorithm. It uses modified packing steps and always maintains a ρ-convex combi- nation satisfying the conditions of Lemma 5. As usual, the core algorithm HbM (Algorithm 3) operates on solutions where the incidence matrix A is of full column rank.

Algorithm 3 HbM(V,E,x) // write x as ρ-convex comb. of 0-1 solutions Require: A has its columns linearly independent 1: If x = 0 return the trivial ρ-convex combination λ1 = ρ,x1 = 0. 2: Pick a vertex vb with minimum nonzero degree. 3: Pick e ∈ δ(vb) such that xe is maximized. i 4: Recurse: (x ,λi)i := HbM(V,E \ {e},x|E\{e}). i E 5: Extend each x back to R by setting the e-coordinates to 0. i 6: Modified packing step: pack xe of e into (x ,λi)i and return the result.

Lemma 6. If 0

Proof. The proof is very similar to proof of Lemma 5 (except we have linear independence instead of ex- tremeness) and we therefore re-use its notation and some of the observations therein. Our goal is to show that each modified packing step succeeds. Write Q for v∈e Qv. For the modified packing step to succeed we need λ(Q)+ xe ≤ ρ as before. We will use that λ(Qv) ≤ (1 − xe) for each v, which holds as in Lemma 5. S The first case we will handle is |e| < k. In this case, λ(Q)+ xe ≤ |e|(1 − xe)+ xe ≤ (k − 1)(1 − xe)+ xe ≤ k − 1 ≤ ρ, as needed. So we assume |e| = k. Since Lemma 1 applies to our setting, the degree of v is at most µ. The next case we will handle is xe ≥ 1/µ. In this case, λ(Q)+ xe ≤ k(1 − xe)+ xe = k − (k − 1)xe ≤ k − (k − 1)/µ = ρ (like the proof of Proposition 2). So we may assume xe < 1/µ. b Likewise, by the definition of µ, we may assume x(δ(v)) < 1, since otherwise we fall in to the previous case by our choice of e. Since x(δ(v)) < 1, we can get an exact expression forbQvb more specific than that given in the proof of Lemma 5. All solutions xi in the ρ-convex combination have degree 0 or 1 at v, and the latter are the ones ′ in Qvb (blockingb e at v), and so λ(Qvb) = (Ax )vb = (Ax)vb − xe = x(δ(v)) − xe. This complements the upper bounds λ(Qv) ≤ 1 − xe that hold for all other v ∈ e with v 6= v. This lets us bound the amount of room needed for the modifiedb packing step: b b xe + λ(Q) ≤ xe + x(δ(v)) − xe + (k − 1)(1 − xe)

≤ µxe + (k − 1)(1 − xe)= k − 1 + (µ − k + 1)xe ≤ k − 1 + (µb− k + 1)/µ = k − (k − 1)/µ = ρ

1 where the middle inequality used x(δ(v)) ≤ µxe and the last used xe < µ . To complete the proofs of Theorems 1 and 2, we yet again use the approach of starting with an extreme b point solution and fixing its integer part (like Observation 1), recursing only on the residual b-matching problem, which has linearly independent rows and 0

5 Application: Bounded-Color k-Hypergraph b-Matching

We observe that improved approximations for the bounded-color k-hypergraph b-matching problem, which is defined above Corollary 1, follow directly from our results. The specialization of this problem for the case

9 of matchings in graphs was very recently introduced by Stamoulis [27], who gave a 2-approximation (note that Stamoulis had considered only matchings and not b-matchings). This independent result also leverages a variant of iterated packing. We give a k-approximation for the general case of bounded-color k-hypergraph b-matching, and thus extend the above result to hypergraphs as well as b-matchings. Stamoulis observed that bounded-color matching is a special case of 3-hypergraph b-matching: for each color class Ei, add a new vertex ci with capacity wi. Now replace each edge {u, v} with a hyperedge {ci,u,v}. This precisely models the bounded-color matching problem. An analogous reduction shows that bounded- color k-hypergraph b-matching is a special case of standard (k + 1)-hypergraph b-matching. To obtain our approximation, we simply observe that these special instances are bipartite, as the set U consisting of all the ci vertices intersects every hyperedge exactly once. This gives us a k-approximation since the instance under consideration is a (k + 1)-hypergraph.

6 Application: Allocations

We will take advantage of the Lavi-Swamy framework [21], which is a fractional version of the well-known Vickrey-Clarke-Groves (VCG) mechanism. We cannot directly use VCG in this setting, because one of the steps in VCG is to compute the allocation which maximizes the total utility of all players, and this problem is NP-complete in our setting for t ≥ 2, by a reduction from 3-dimensional matching. The main result of Lavi and Swamy is that once we have an LP-relative ρ-approximation algorithm with respect to the natural LP, we can get a truthful-in-expectation mechanism, which also maximizes the expected overall utility within a factor of ρ. Minimizing this factor means we are coming closer to a VCG-like mechanism, whereas allocating everyone the empty set is truthful but a bad approximation. i First we define the natural LP relaxation for the allocation problem. Let xS be a fractional indicator variable indicating whether player i will win exactly the set S of items. Then the LP requires that each player i wins one set of items, and that each item is allocated at most once, fractionally. Write vS as the valuation of player i for set S. Altogether the fractional allocation LP is:

i i i i max xS vS :0 ≤ x ≤ 1; ∀i ∈ [n]: xS = 1; ∀s ∈ [m]: xS ≤ 1. (A) i,S S i S s∈S X X X X: We assume the input to the mechanism is an explicit list from each bidder, consisting of their valuation for each set upon which they wish to put a positive bid. The number of variables and constraints in the LP is polynomial in the number of such bids. Although for constant k, any reasonable bid language or oracle can be used, since the number of sets of size < k is polynomial and we can convert everything to an explicit list. Definition 3. An ρ-approximate truthful-in-expectation mechanism for the allocation problem is a random- ized algorithm of the following form. It takes the values v as inputs; its outputs are a valid allocation of items to players together with prices pi charged to each player i. It has the following two properties. First, where i i S(i) denotes the set of items allocated to player i, we have i vS(i) is at least i vT (i)/ρ for every valid allocation T . Second, for every fixed v−i, a player who gives insincere valuations vi as their input, resulting P P in random variables p, S compared to the original ones p,S, does not increase their expected net utility: i i b E[v b − pi] ≤ E[vS(i) − pi]. b b S(i) Moreover, 0 ≤ E[p ] ≤ E[vi ] for all i. i S(i) b Theorem 4 (Lavi-Swamy [21]). Given a polynomial-time LP-relative ρ-approximation algorithm for an allocation problem, we can obtain a polynomial-time ρ-approximate truthful-in-expectation mechanism. However, the allocation problem here is precisely bipartite k-hypergraph matching: for each bidder and each set of items they could win, create a set out of them all together, and this set has size at most 1+ k − 1= k; and each such hyperedge contains exactly one bidder, so the hypergraph is indeed bipartite. So our bipartite extension of the Chan-Lau theorem (Section 2) applies and we are done. The LP-relative property is essential; the non-LP relative local search approach from [13] cannot be used with [21].

10 7 Local Ratio and k-Hypergraph Demand Matching

We recommend [4,6,7] for background on the local ratio method, including its relationship with the primal- dual method. The heart of the local ratio approach is the following lemma:

Lemma 7 (Local ratio lemma). Let xOP T be the (unknown) optimal integral solution. If wi · xLR ≥ wi · xOP T for all i, and w = i wi, then w · xLR ≥ w · xOP T , i.e. xLR is α-approximately optimal. Compared with fractional localP ratio, we do not start by solving an LP, which is faster. But, we cannot use x∗ to guide the algorithm — we have to ensure an oblivious approximation guarantee that holds against the unknown optimal solution. In this section we briefly outline a reinterpretation of the 2k-approximation for k-hypergraph demand matching from [23] as a local ratio algorithm. Compared with [23], the new algorithm will be both simpler and faster (as we solve no LPs). The inspiration for this simplified algorithm is a connection between local ratio algorithm and iterated packing elucidated by Bar-Yehuda et al. [5, p. 12]. As before, let A be the incidence matrix, and let A[d] be the same matrix but with the column for each e having its entries multiplied by de. Then an ILP formulation for the hypergraph demand matching problem is to find an integral x maximizing wx subject to A[d]x ≤ b and c ≥ x ≥ 0. We will assume that de ≤ bv whenever v ∈ e. This is without loss of generality for the purposes of approximation, while for bounding the integrality gap this no-clipping assumption is needed to even get a constant upper bound (even if k = 1, a.k.a. knapsack). We use the same basic ideas used in [23] but arranged differently. The crux in our case is to show that for every instance, there is a hyperedge e and a weight function satisfying that any feasible solution is either 2k-approximately optimal or has room for e to be added. With this (Lemma 8) and using the local ratio lemma, we can show that Algorithm 4 is a 2k-approximation algorithm.

Lemma 8. Let e be the hyperedge so that de is minimal. Define a weight function w on all hyperedges by we =1, and for all other f, df wf := . b (2) max{bv − de, de} b v∈e∩f X Then (i) every feasible solution (whetherb or not it contains e) has value at most 2k under w, (ii) we ≥ 1, and (iii) any feasible subset of E \{e} to which e cannot be added has weight at least 1 under w. b b b Algorithm 4 HDM(V,E,d,b,w) // for hypergraph demand matching 1: Pick e ∈ E such that de is minimum, or return ∅ if E = ∅. E 2: Define a new weight function wb ∈ R via wbe = 1 and (2) for f 6= e. ′ ′ 3: Let wewb be its scalar multiple by we, and w := w − wewb. // note we = 0 ′ ′ ′ 4: Define E := {e ∈ E | we > 0}. // note e 6∈ E ′ ′ ′ 5: Recurse: F := HDM(V,E ,d,b,w |E′ ). 6: If F ′ ∪ {e} is feasible define F := F ′ ∪ {e}, else define F := F ′. 7: Return F.

References

1. R. Aharoni and O. Kessler. On a possible extension of hall’s theorem to bipartite hypergraphs. Discrete Mathe- matics, 84(3):309 – 313, 1990. 2. E. M. Arkin and R. Hassin. On local search for weighted k-set packing. Math. Oper. Res., 23(3):640–648, Mar. 1998.

11 3. N. Bansal, N. Korula, V. Nagarajan, and A. Srinivasan. On k-column sparse packing programs. In Proc. 14th IPCO, pages 369–382, 2010. 4. R. Bar-Yehuda, K. Bendel, A. Freund, and D. Rawitz. Local ratio: A unified framework for approximation algorithms. ACM Comp. Surv., 36(4):422–463, Dec. 2004. 5. R. Bar-Yehuda, M. M. Halld´orsson, J. Naor, H. Shachnai, and I. Shapira. Scheduling split intervals. SIAM J. Comput., 36(1):1–15, 2006. 6. R. Bar-Yehuda and D. Rawitz. On the equivalence between the primal-dual schema and the local ratio technique. SIAM J. Disc. Math., 19(3):762–797, 2005. 7. R. Bar-Yehuda and D. Rawitz. A tale of two methods. Theoretical Computer Science, pages 196–217, 2006. 8. P. Berman. A d/2 approximation for maximum weight independent set in d-claw free graphs. In Proc. 7th SWAT, pages 214–219, 2000. 9. R. D. Carr and S. Vempala. Randomized metarounding. Random Struct. Algorithms, 20(3):343–352, 2002. 10. Y. Chan and L. Lau. On linear and semidefinite programming relaxations for hypergraph matching. Math. Program., pages 1–26, 2011. 11. C. Chekuri, M. Mydlarz, and F. B. Shepherd. Multicommodity demand flow in a tree and packing integer programs. ACM Transactions on Algorithms, 3(3), 2007. 12. U. Feige and M. Singh. Edge coloring and decompositions of weighted graphs. In Proc. ESA, pages 405–416, 2008. 13. M. Feldman, J. Naor, R. Schwartz, and J. Ward. Improved approximations for k-exchange systems. In Proc. 19th ESA, pages 784–798, 2011. 14. Z. F¨uredi. Maximum degree and fractional matchings in uniform hypergraphs. Combinatorica, 1:155–162, 1981. 15. Z. F¨uredi, J. Kahn, and P. D. Seymour. On the fractional matching polytope of a hypergraph. Combinatorica, 13(2):167–180, 1993. 16. E. Hazan, S. Safra, and O. Schwartz. On the complexity of approximating k-set packing. Computational Com- plexity, 15(1):20–39, 2006. 17. C. A. J. Hurkens and A. Schrijver. On the size of systems of sets every t of which have an SDR, with an application to the worst-case ratio of heuristics for packing problems. SIAM J. Discrete Math., 2(1):68–72, 1989. 18. V. Kann. Maximum bounded 3-dimensional matching is MAX SNP-complete. Information Processing Letters, 37(1):27–35, 1991. 19. C. Koufogiannakis and N. E. Young. Distributed fractional packing and maximum weighted b-matching via tail-recursive duality. In Proc. 23rd DISC, pages 221–238, 2009. 20. P. Krysta. Greedy approximation via duality for packing, combinatorial auctions and routing. In Proc. 30th MFCS, pages 615–627, 2005. 21. R. Lavi and C. Swamy. Truthful and near-optimal mechanism design via linear programming. In Proc. 46th FOCS, pages 595–604, 2005. 22. N. Nisan, T. Roughgarden, E.´ Tardos, and V. V. Vazirani. Algorithmic Game Theory. Cambridge University Press, New York, NY, USA, 2007. 23. O. Parekh. Iterative packing for demand and hypergraph matching. In Proc. 15th IPCO, pages 349–361, 2011. 24. Y. Person and M. Schacht. Almost all hypergraphs without fano planes are bipartite. In Proceedings of the twentieth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’09, pages 217–226, Philadelphia, PA, USA, 2009. Society for Industrial and Applied Mathematics. 25. A. Schrijver. Combinatorial optimization. Springer, New York, 2003. 26. F. B. Shepherd and A. Vetta. The demand-matching problem. Mathematics of Operations Research, 32(3):563– 578, August 2007. 27. G. Stamoulis. Approximation algorithms for bounded color matchings via convex decompositions. In E. Csuhaj- Varj´u, M. Dietzfelbinger, and Z. Esik,´ editors, Mathematical Foundations of Computer Science 2014 - 39th International Symposium, MFCS 2014, Budapest, Hungary, August 25-29, 2014. Proceedings, Part II, volume 8635 of Lecture Notes in Computer Science, pages 625–636. Springer, 2014. 28. W. Tutte. A short proof of the factor theorem for finite graphs. Canadian Journal of Mathematics, 6:347–352, 1954.

12